基于深度学习的单目标快速跟踪方法  

Fast Single Target Tracking Method Based on Deep Learning

作　　者：杨思诚 YANG Sicheng(School of Information Science&Engineering,Yunnan University,Kunming,Yunnan 650500,China)

机构地区：[1]云南大学信息学院,云南昆明650500

出　　处：《自动化应用》2025年第6期57-60,63,共5页Automation Application

摘　　要：在学术领域,单目标跟踪问题长期占据研究前沿,尤其是在体育足球分析场景中,其复杂多变的特性(如目标尺度动态变化、复杂背景干扰以及目标高速运动等)成为该领域内的主要挑战。鉴于此,提出了一种创新的基于深度学习的足球单目标跟踪算法。首先,通过深度优化核相关滤波技术,解决了目标跟踪中的分类难题,显著提升了分类精度与鲁棒性。然后,针对ATOM算法中的核心——目标估计模块,实施了网络结构的精细化调整,并引入了ResNet-50这一高效深度神经网络进行再训练,以增强模型对复杂运动模式的捕捉能力,同时保持计算效率。为验证所提算法的性能,在大规模的GOT-10K足球视频数据集上进行广泛的实验评估。实验结果表明,该算法不仅能准确追踪足球运动目标,即使在目标快速移动、尺度变化及背景复杂等极端情况下也能表现出色,而且其运行速度达到70fps,充分证明了其在实时性应用中的潜力。综上,所提出的基于深度学习的足球单目标跟踪方法在解决体育足球领域单目标跟踪问题上展现出了显著的优势与实用性。In the academic field,single target tracking has long been at the forefront of research,especially in sports and football analysis scenarios,where its complex and variable characteristics(such as dynamic changes in target scale,complex background interference,and high-speed target movement)pose the main challenges in this field.In view of this,an innovative football single target tracking algorithm based on deep learning is proposed.Firstly,by deeply optimizing kernel correlation filtering technology,the classification problem in target tracking has been solved,significantly improving classification accuracy and robustness.Then,focusing on the core of the ATOM algorithm-the target estimation module,a refined adjustment of the network structure was implemented,and ResNet-50,an efficient deep neural network,was introduced for retraining to enhance the model's ability to capture complex motion patterns while maintaining computational efficiency.To validate the performance of the proposed algorithm,extensive experimental evaluations were conducted on a large-scale GOT-10K football video dataset.The experimental results show that the algorithm can not only accurately track football targets,but also perform well even in extreme situations such as rapid target movement,scale changes,and complex backgrounds.Moreover,its running speed reaches 70 fps,fully demonstrating its potential in real-time applications.In summary,the proposed deep learning based football single target tracking method has demonstrated significant advantages and practicality in solving the problem of single target tracking in the field of sports football.

关 键 词：单目标跟踪 深度学习 ATOM算法 

分 类 号：TP273.4[自动化与计算机技术—检测技术与自动化装置]